Architecture Overview · Park Whisperer Platform

Automated Social Content Factory

A 3-pass content generation pipeline that pre-fetches all operational data before any LLM call, then uses Claude Haiku for agentic data synthesis and Claude Sonnet for platform-specific formatting — producing blog posts, social captions, and Reels scripts from live park data.

● Production · 10 Pipeline TypesClaude Haiku + Sonnet0 Hallucinated Data

Click any component to explore it in the full case study.

Input A
Scheduled Pipeline Trigger
10 pipeline types: operations_bulletin, weather_report, morning_briefing, rope_drop, crowd_update, and more. Each trigger pre-defines knowledge_types, tool set, and output schema.
Scheduled · Azure Functions
Input B
Pre-Fetch Knowledge Pack
All knowledge queries and real-time tool calls resolved BEFORE the LLM sees any tokens. KPI fact pack: crowd_score, top_10_waits, weather_conditions, sellout_events.
Pre-fetch · Parallel tools
3-Pass Content Generation Pipeline
KPI Gate
Data Validation
Validates crowd_score present, operational_date matches. Blocks generation if KPI data missing.
›
Pass 1
Data Synthesis
Claude Haiku agentic loop — 30+ tool calls via ThreadPoolExecutor. Assembles structured summary from live operational data.
›
Pass 2
Platform Format
Claude Sonnet — formats for Facebook, Instagram, Threads. Per-platform character limits, tone, hashtag strategy.
›
Pass 3
Reels Script
HOOK + SEGMENT_1 + CTA + NARRATION. 36–42 words enforced. Feeds directly into AI Video Team pipeline.
🛡️
Anti-Hallucination Architecture
Pre-fetch layer ensures all data is real before generation starts. Tool failures return "No data available" — never hallucinate. KPI gate blocks publishing without operational data. Gate reason audit tracks every failure.
Pre-fetch completeness gateKPI gate (no data = no publish)Tool failure = "No data available"Gate reason audit trailParallel tool execution tracking0 hallucinated data points
Output A
Social Content Artifact
Per-platform sections: Facebook post, Instagram caption, Threads post. Platform-native formatting, character limits enforced, hashtag strategy per pipeline type.
Facebook · Instagram · Threads
Output B
Blog Post
Long-form article with operational data grounding. SEO-structured, park-voice consistent, all claims backed by real wait time / sellout / weather data.
Long-form · Data-grounded
Output C
Reels Script → Video Queue
36–42 word NARRATION block triggers AI Video Team pipeline. HOOK + SEGMENT + CTA structure feeds ElevenLabs TTS and FFmpeg render.
36–42 words · Video trigger
10
Pipeline types
30+
LangChain tools
3
Content passes
0
Hallucinated data
StackPythonAzure FunctionsLangChainClaude Haiku (Pass 1)Claude Sonnet (Pass 2–3)PostgreSQL RAGThreadPoolExecutorCosmos DB